Which Business Processes to Automate with AI (And Which to Leave Alone)

By Simeon Matheka, Founder & Creative Director · Published 2026-09-07 · Updated 2026-09-07 · 17 min read

A scored catalog for AI-based business automation: volume, unstructured input, and the cost of a wrong send. Which loops get a model, which stay rules, and which stay human. First-party examples from intake we actually run.

You do not need another definition of AI business process automation. You need a rule for which loops get a model this month, which stay a boring workflow, and which stay a person with a checklist.

The vocabulary lives in what AI business process automation is. This page is the filter: score the process before you buy a seat.

The score: volume, mess, blast radius

AI-based business automation earns its keep on a loop that is frequent, arrives messy, and is cheap to undo. Miss any one of those and a rules workflow (or a human) is cheaper.

If you cannot write the standard operating procedure (SOP), stop. Automating a fuzzy intake just creates faster mess. The SOPs before automation playbook is the pre-filter. This catalog assumes the SOP exists.

flowchart TD
    Proc["Candidate process"] --> Sop{"SOP written<br/>happy path + exceptions?"}
    Sop -->|No| Stop["Leave human<br/>write the SOP first"]
    Sop -->|Yes| Shape{"Input mostly<br/>unstructured?"}
    Shape -->|No| Rules["Rules workflow<br/>no model"]
    Shape -->|Yes| Blast{"Wrong send<br/>cheap to undo?"}
    Blast -->|No| Hitl["Model drafts<br/>human-in-the-loop sends"]
    Blast -->|Yes| Vol{"Happens often<br/>enough to pay back?"}
    Vol -->|No| Skip["Skip this month"]
    Vol -->|Yes| Ai["Model on the messy step<br/>rules on the write"]

Catalog: rules, AI, or stop

Use this as a pick list, not a mandate. “AI” here means a large language model (LLM) or classifier on one step. The rest of the process stays deterministic: validate, store, notify, audit.

Usually a model (with a schema)

• Inbound lead triage: Website form, WhatsApp, or email arrives as prose. The model extracts intent, budget band, and urgency into structured fields. Rules then write the Customer Relationship Management (CRM) row and ping Slack. We run this shape for studio intake.
• Support ticket routing: Classify topic and tone. Do not auto-reply on the first version. Route, then let a human send.
• Document field extraction: Invoices, briefs, contracts as PDFs. The model proposes fields. A person confirms amounts and names before anything posts to accounting.
• Meeting or call notes to a CRM row (after the first extract): Draft the summary and next steps. A person clicks approve. Never let the model invent a deal value.
• Internal first drafts: Job posts, FAQ answers, outline for a page. Human publishes. Low blast radius if it stays internal until review.

Usually rules only

• Form POST to a database: Validate at the edge, insert the row, notify. No model. Our contact stack does this with Supabase and edge functions.
• Invoice paid → provision access: Webhook, signature check, plan lookup, write. A model has nothing to interpret.
• Appointment booked → calendar + reminder: Structured event. Rules and a scheduler. Save the tokens.
• Failed payment retry: Stripe (or your processor) already has the state machine. Do not wrap it in a chatty agent.

Stop, or human-in-the-loop only

• Anything that emails the client: Draft in a queue. A person hits send. The human-in-the-loop review article is the queue, retry, and kill switch.
• Money movement and refunds above a small cap: The model can propose. Policy code decides. A person approves over the cap.
• Legal, HR firing, medical, or “the founder decides”: Not a workflow. A person with context.
• A process nobody can draw: You will encode folklore. It will fire at 2 a.m. and you will not know why.

What we actually run (not a vendor catalog)

Three loops on work we ship. Score them the same way you score yours.

The lead-routing architecture (webhooks, schema drift, retries, 0% silent drops) is the n8n and Supabase lead routing case study. The form is static HTML on Cloudflare. The POST is not. Do not bolt a PHP mailer onto either loop.

Department cheat sheet

Same score, different desks. If two rows conflict, blast radius wins.

Thirty-day sprint: one process

Do not automate the company. Automate one scored loop. If you cannot name the owner and the undo, you are not ready.

1. Days 1–3, pick: List ten loops. Score volume, mess, blast radius, SOP. Take the highest score that does not email the client.
2. Days 4–7, freeze the SOP: Happy path, exceptions, what “done” looks like, who gets the Slack. If this takes more than a page, the process is not a candidate.
3. Days 8–14, rules first: Trigger, validate, store, notify. No model. Prove the pipes. Measure how many rows per week.
4. Days 15–21, model on one step: Extract or classify into a JSON schema. Log raw in and structured out. Do not send yet.
5. Days 22–30, gate: Confidence threshold, review queue for misses, cost cap, kill switch. Only then consider a client-facing draft.

Platform pick (n8n vs Zapier vs Make) is a later article. The step-by-step how to automate guide is the build sequence once this filter has a winner.

Failure modes when the score was wrong

• You modeled a structured webhook: Token bill for a switch statement. Delete the model. Keep the workflow.
• You skipped the SOP: Two weeks in, nobody agrees what “qualified” means. The classifier oscillates. Write the page. Then retrain the prompt.
• You let the model send: One hallucinated price or name and you spend the savings on an apology. Queue first, forever, for anything external.
• You automated ten loops at once: No baseline, no owner, no undo. Pick one. The complete guide is the landscape after you have a win, not a shopping list for week one.
• You cloned a definition page: Another “what is” article does not help the operator pick a loop. Publish the score, the catalog, and one process you actually run.

Put the filter on the wall

Rules for clean events. A model for messy text. A person for send, money, and anything you cannot undo. If a process does not score on volume, mess, and blast radius, it is not this month’s project.

Print the AI process scorecard and fill three loops before you open a builder. If you want help scoring a real intake or delivery loop, start a conversation. Bring the SOP, or the fact that you do not have one. That is the useful first meeting.

Tags: AI-based business automation, business process automation, AI automation, process selection, n8n, human-in-the-loop, SOPs, workflow automation